diff --git "a/phowhisper-base/TextDecoder.mlmodelc/model.mil" "b/phowhisper-base/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/phowhisper-base/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,1115 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}, {"coremltools-component-torch", "2.1.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})] +{ + func main(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { + tensor var_28_axis_0 = const()[name = tensor("op_28_axis_0"), val = tensor(0)]; + tensor var_28_batch_dims_0 = const()[name = tensor("op_28_batch_dims_0"), val = tensor(0)]; + tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_28_cast_fp16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_28_cast_fp16")]; + tensor var_32_axis_0 = const()[name = tensor("op_32_axis_0"), val = tensor(0)]; + tensor var_32_batch_dims_0 = const()[name = tensor("op_32_batch_dims_0"), val = tensor(0)]; + tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53109888)))]; + tensor var_32_cast_fp16 = gather(axis = var_32_axis_0, batch_dims = var_32_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_32_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_28_cast_fp16, y = var_32_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_46_axes_0 = const()[name = tensor("op_46_axes_0"), val = tensor([2])]; + tensor var_46_cast_fp16 = expand_dims(axes = var_46_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_46_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_46_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([512, 512, 512, 512, 512, 512])]; + tensor var_51_axis_0 = const()[name = tensor("op_51_axis_0"), val = tensor(1)]; + tensor var_51_cast_fp16_0, tensor var_51_cast_fp16_1, tensor var_51_cast_fp16_2, tensor var_51_cast_fp16_3, tensor var_51_cast_fp16_4, tensor var_51_cast_fp16_5 = split(axis = var_51_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_51_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([512, 512, 512, 512, 512, 512])]; + tensor var_60_axis_0 = const()[name = tensor("op_60_axis_0"), val = tensor(1)]; + tensor var_60_cast_fp16_0, tensor var_60_cast_fp16_1, tensor var_60_cast_fp16_2, tensor var_60_cast_fp16_3, tensor var_60_cast_fp16_4, tensor var_60_cast_fp16_5 = split(axis = var_60_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_60_cast_fp16")]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(3)]; + tensor var_79 = const()[name = tensor("op_79"), val = tensor(1)]; + tensor var_80 = const()[name = tensor("op_80"), val = tensor(true)]; + tensor var_92 = const()[name = tensor("op_92"), val = tensor([1])]; + tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_92, keep_dims = var_80, x = inputs_1_cast_fp16)[name = tensor("channels_mean_1_cast_fp16")]; + tensor zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor("zero_mean_1_cast_fp16")]; + tensor zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor("zero_mean_sq_1_cast_fp16")]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1])]; + tensor var_97_cast_fp16 = reduce_mean(axes = var_96, keep_dims = var_80, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_97_cast_fp16")]; + tensor var_98_to_fp16 = const()[name = tensor("op_98_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_99_cast_fp16 = add(x = var_97_cast_fp16, y = var_98_to_fp16)[name = tensor("op_99_cast_fp16")]; + tensor denom_1_epsilon_0_to_fp16 = const()[name = tensor("denom_1_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_99_cast_fp16)[name = tensor("denom_1_cast_fp16")]; + tensor out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53568704)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53569792)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53570880)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53571968)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor var_114 = const()[name = tensor("op_114"), val = tensor([1, 1])]; + tensor var_116 = const()[name = tensor("op_116"), val = tensor([1, 1])]; + tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("custom")]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53573056)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54097408)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_116, groups = var_79, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_114, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor var_120 = const()[name = tensor("op_120"), val = tensor([1, 1])]; + tensor var_122 = const()[name = tensor("op_122"), val = tensor([1, 1])]; + tensor current_key_1_pad_type_0 = const()[name = tensor("current_key_1_pad_type_0"), val = tensor("custom")]; + tensor current_key_1_pad_0 = const()[name = tensor("current_key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54098496)))]; + tensor current_key_1_cast_fp16 = conv(dilations = var_122, groups = var_79, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_120, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor var_127 = const()[name = tensor("op_127"), val = tensor([1, 1])]; + tensor var_129 = const()[name = tensor("op_129"), val = tensor([1, 1])]; + tensor current_value_1_pad_type_0 = const()[name = tensor("current_value_1_pad_type_0"), val = tensor("custom")]; + tensor current_value_1_pad_0 = const()[name = tensor("current_value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54622848)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55147200)))]; + tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_129, groups = var_79, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_127, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_133_axes_0 = const()[name = tensor("op_133_axes_0"), val = tensor([1])]; + tensor var_133_cast_fp16 = expand_dims(axes = var_133_axes_0, x = kv_cache_update_mask)[name = tensor("op_133_cast_fp16")]; + tensor var_134_axes_0 = const()[name = tensor("op_134_axes_0"), val = tensor([2])]; + tensor var_134_cast_fp16 = expand_dims(axes = var_134_axes_0, x = var_133_cast_fp16)[name = tensor("op_134_cast_fp16")]; + tensor var_136_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_136_cast_fp16")]; + tensor var_73_to_fp16 = const()[name = tensor("op_73_to_fp16"), val = tensor(0x1p+0)]; + tensor var_137_cast_fp16 = sub(x = var_73_to_fp16, y = var_134_cast_fp16)[name = tensor("op_137_cast_fp16")]; + tensor var_138_cast_fp16 = mul(x = var_51_cast_fp16_0, y = var_137_cast_fp16)[name = tensor("op_138_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_136_cast_fp16, y = var_138_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_140_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_140_cast_fp16")]; + tensor var_142_cast_fp16 = mul(x = var_60_cast_fp16_0, y = var_137_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_140_cast_fp16, y = var_142_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_145 = const()[name = tensor("op_145"), val = tensor([1, 8, 64, -1])]; + tensor var_146_cast_fp16 = reshape(shape = var_145, x = query_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor var_147_to_fp16 = const()[name = tensor("op_147_to_fp16"), val = tensor(0x1p-3)]; + tensor var_148_cast_fp16 = mul(x = var_146_cast_fp16, y = var_147_to_fp16)[name = tensor("op_148_cast_fp16")]; + tensor var_149 = const()[name = tensor("op_149"), val = tensor([1, 8, 64, -1])]; + tensor var_150_cast_fp16 = reshape(shape = var_149, x = key_1_cast_fp16)[name = tensor("op_150_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_148_cast_fp16, y = var_150_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_154_axes_0 = const()[name = tensor("op_154_axes_0"), val = tensor([1])]; + tensor var_154_cast_fp16 = expand_dims(axes = var_154_axes_0, x = decoder_key_padding_mask)[name = tensor("op_154_cast_fp16")]; + tensor var_155_axes_0 = const()[name = tensor("op_155_axes_0"), val = tensor([2])]; + tensor var_155_cast_fp16 = expand_dims(axes = var_155_axes_0, x = var_154_cast_fp16)[name = tensor("op_155_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_155_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_158_cast_fp16 = softmax(axis = var_72, x = mh_w_3_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_159 = const()[name = tensor("op_159"), val = tensor([1, 8, 64, -1])]; + tensor var_160_cast_fp16 = reshape(shape = var_159, x = value_1_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_160_cast_fp16, y = var_158_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 512, 1, -1])]; + tensor input_1_cast_fp16 = reshape(shape = var_163, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_167 = const()[name = tensor("op_167"), val = tensor([1, 1])]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor([1, 1])]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("custom")]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55148288)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55672640)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_169, groups = var_79, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_167, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor var_179 = const()[name = tensor("op_179"), val = tensor([1])]; + tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_179, keep_dims = var_80, x = inputs_3_cast_fp16)[name = tensor("channels_mean_3_cast_fp16")]; + tensor zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor("zero_mean_3_cast_fp16")]; + tensor zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor("zero_mean_sq_3_cast_fp16")]; + tensor var_183 = const()[name = tensor("op_183"), val = tensor([1])]; + tensor var_184_cast_fp16 = reduce_mean(axes = var_183, keep_dims = var_80, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_184_cast_fp16")]; + tensor var_185_to_fp16 = const()[name = tensor("op_185_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_186_cast_fp16 = add(x = var_184_cast_fp16, y = var_185_to_fp16)[name = tensor("op_186_cast_fp16")]; + tensor denom_3_epsilon_0_to_fp16 = const()[name = tensor("denom_3_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_186_cast_fp16)[name = tensor("denom_3_cast_fp16")]; + tensor out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55673728)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55674816)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 1])]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1])]; + tensor query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("custom")]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(55675904)))]; + tensor layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56200256)))]; + tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_203, groups = var_79, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_201, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_207 = const()[name = tensor("op_207"), val = tensor([1, 1])]; + tensor var_209 = const()[name = tensor("op_209"), val = tensor([1, 1])]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("custom")]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56201344)))]; + tensor key_3_cast_fp16 = conv(dilations = var_209, groups = var_79, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_207, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor([1, 1])]; + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 1])]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("custom")]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56725696)))]; + tensor layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57250048)))]; + tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_216, groups = var_79, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_214, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; + tensor var_220 = const()[name = tensor("op_220"), val = tensor([1, 8, 64, -1])]; + tensor var_221_cast_fp16 = reshape(shape = var_220, x = query_3_cast_fp16)[name = tensor("op_221_cast_fp16")]; + tensor var_222_to_fp16 = const()[name = tensor("op_222_to_fp16"), val = tensor(0x1p-3)]; + tensor var_223_cast_fp16 = mul(x = var_221_cast_fp16, y = var_222_to_fp16)[name = tensor("op_223_cast_fp16")]; + tensor var_224 = const()[name = tensor("op_224"), val = tensor([1, 8, 64, -1])]; + tensor var_225_cast_fp16 = reshape(shape = var_224, x = key_3_cast_fp16)[name = tensor("op_225_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_223_cast_fp16, y = var_225_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_72, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 8, 64, -1])]; + tensor var_230_cast_fp16 = reshape(shape = var_229, x = value_3_cast_fp16)[name = tensor("op_230_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_230_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 512, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = var_233, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 1])]; + tensor var_239 = const()[name = tensor("op_239"), val = tensor([1, 1])]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("custom")]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57251136)))]; + tensor layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57775488)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_239, groups = var_79, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_237, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_245 = const()[name = tensor("op_245"), val = tensor([1])]; + tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_245, keep_dims = var_80, x = inputs_5_cast_fp16)[name = tensor("channels_mean_5_cast_fp16")]; + tensor zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor("zero_mean_5_cast_fp16")]; + tensor zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor("zero_mean_sq_5_cast_fp16")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor([1])]; + tensor var_250_cast_fp16 = reduce_mean(axes = var_249, keep_dims = var_80, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_to_fp16 = const()[name = tensor("op_251_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_252_cast_fp16 = add(x = var_250_cast_fp16, y = var_251_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor denom_5_epsilon_0_to_fp16 = const()[name = tensor("denom_5_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_252_cast_fp16)[name = tensor("denom_5_cast_fp16")]; + tensor out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_5_gamma_0_to_fp16 = const()[name = tensor("input_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57776576)))]; + tensor input_5_beta_0_to_fp16 = const()[name = tensor("input_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777664)))]; + tensor input_5_epsilon_0_to_fp16 = const()[name = tensor("input_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor var_263 = const()[name = tensor("op_263"), val = tensor([1, 1])]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1])]; + tensor input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("custom")]; + tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc1_weight_to_fp16 = const()[name = tensor("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57778752)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59875968)))]; + tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_265, groups = var_79, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_263, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("EXACT")]; + tensor input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor var_271 = const()[name = tensor("op_271"), val = tensor([1, 1])]; + tensor var_273 = const()[name = tensor("op_273"), val = tensor([1, 1])]; + tensor hidden_states_3_pad_type_0 = const()[name = tensor("hidden_states_3_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_3_pad_0 = const()[name = tensor("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc2_weight_to_fp16 = const()[name = tensor("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59880128)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61977344)))]; + tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_273, groups = var_79, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_271, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor(3)]; + tensor var_293 = const()[name = tensor("op_293"), val = tensor(1)]; + tensor var_294 = const()[name = tensor("op_294"), val = tensor(true)]; + tensor var_306 = const()[name = tensor("op_306"), val = tensor([1])]; + tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_306, keep_dims = var_294, x = inputs_7_cast_fp16)[name = tensor("channels_mean_7_cast_fp16")]; + tensor zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor("zero_mean_7_cast_fp16")]; + tensor zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor("zero_mean_sq_7_cast_fp16")]; + tensor var_310 = const()[name = tensor("op_310"), val = tensor([1])]; + tensor var_311_cast_fp16 = reduce_mean(axes = var_310, keep_dims = var_294, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_311_cast_fp16")]; + tensor var_312_to_fp16 = const()[name = tensor("op_312_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_313_cast_fp16 = add(x = var_311_cast_fp16, y = var_312_to_fp16)[name = tensor("op_313_cast_fp16")]; + tensor denom_7_epsilon_0_to_fp16 = const()[name = tensor("denom_7_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_313_cast_fp16)[name = tensor("denom_7_cast_fp16")]; + tensor out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61978432)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61979520)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor var_328 = const()[name = tensor("op_328"), val = tensor([1, 1])]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 1])]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("custom")]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61980608)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62504960)))]; + tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_330, groups = var_293, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_328, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_334 = const()[name = tensor("op_334"), val = tensor([1, 1])]; + tensor var_336 = const()[name = tensor("op_336"), val = tensor([1, 1])]; + tensor current_key_3_pad_type_0 = const()[name = tensor("current_key_3_pad_type_0"), val = tensor("custom")]; + tensor current_key_3_pad_0 = const()[name = tensor("current_key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(62506048)))]; + tensor current_key_3_cast_fp16 = conv(dilations = var_336, groups = var_293, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = var_334, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor var_341 = const()[name = tensor("op_341"), val = tensor([1, 1])]; + tensor var_343 = const()[name = tensor("op_343"), val = tensor([1, 1])]; + tensor current_value_3_pad_type_0 = const()[name = tensor("current_value_3_pad_type_0"), val = tensor("custom")]; + tensor current_value_3_pad_0 = const()[name = tensor("current_value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63030400)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63554752)))]; + tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_343, groups = var_293, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = var_341, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_350_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_350_cast_fp16")]; + tensor var_352_cast_fp16 = mul(x = var_51_cast_fp16_1, y = var_137_cast_fp16)[name = tensor("op_352_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_350_cast_fp16, y = var_352_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_354_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_356_cast_fp16 = mul(x = var_60_cast_fp16_1, y = var_137_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_354_cast_fp16, y = var_356_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_359 = const()[name = tensor("op_359"), val = tensor([1, 8, 64, -1])]; + tensor var_360_cast_fp16 = reshape(shape = var_359, x = query_5_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_361_to_fp16 = const()[name = tensor("op_361_to_fp16"), val = tensor(0x1p-3)]; + tensor var_362_cast_fp16 = mul(x = var_360_cast_fp16, y = var_361_to_fp16)[name = tensor("op_362_cast_fp16")]; + tensor var_363 = const()[name = tensor("op_363"), val = tensor([1, 8, 64, -1])]; + tensor var_364_cast_fp16 = reshape(shape = var_363, x = key_5_cast_fp16)[name = tensor("op_364_cast_fp16")]; + tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; + tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; + tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_362_cast_fp16, y = var_364_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_155_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_372_cast_fp16 = softmax(axis = var_286, x = mh_w_9_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_373 = const()[name = tensor("op_373"), val = tensor([1, 8, 64, -1])]; + tensor var_374_cast_fp16 = reshape(shape = var_373, x = value_5_cast_fp16)[name = tensor("op_374_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_374_cast_fp16, y = var_372_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 512, 1, -1])]; + tensor input_11_cast_fp16 = reshape(shape = var_377, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_381 = const()[name = tensor("op_381"), val = tensor([1, 1])]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 1])]; + tensor obj_21_pad_type_0 = const()[name = tensor("obj_21_pad_type_0"), val = tensor("custom")]; + tensor obj_21_pad_0 = const()[name = tensor("obj_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63555840)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64080192)))]; + tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_383, groups = var_293, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = var_381, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1])]; + tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_393, keep_dims = var_294, x = inputs_9_cast_fp16)[name = tensor("channels_mean_9_cast_fp16")]; + tensor zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor("zero_mean_9_cast_fp16")]; + tensor zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor("zero_mean_sq_9_cast_fp16")]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor([1])]; + tensor var_398_cast_fp16 = reduce_mean(axes = var_397, keep_dims = var_294, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_400_cast_fp16 = add(x = var_398_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor denom_9_epsilon_0_to_fp16 = const()[name = tensor("denom_9_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_400_cast_fp16)[name = tensor("denom_9_cast_fp16")]; + tensor out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64081280)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64082368)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor var_415 = const()[name = tensor("op_415"), val = tensor([1, 1])]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 1])]; + tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("custom")]; + tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64083456)))]; + tensor layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64607808)))]; + tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_417, groups = var_293, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = var_415, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_421 = const()[name = tensor("op_421"), val = tensor([1, 1])]; + tensor var_423 = const()[name = tensor("op_423"), val = tensor([1, 1])]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("custom")]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64608896)))]; + tensor key_7_cast_fp16 = conv(dilations = var_423, groups = var_293, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = var_421, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_7_cast_fp16")]; + tensor var_428 = const()[name = tensor("op_428"), val = tensor([1, 1])]; + tensor var_430 = const()[name = tensor("op_430"), val = tensor([1, 1])]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("custom")]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65133248)))]; + tensor layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65657600)))]; + tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_430, groups = var_293, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = var_428, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_7_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 8, 64, -1])]; + tensor var_435_cast_fp16 = reshape(shape = var_434, x = query_7_cast_fp16)[name = tensor("op_435_cast_fp16")]; + tensor var_436_to_fp16 = const()[name = tensor("op_436_to_fp16"), val = tensor(0x1p-3)]; + tensor var_437_cast_fp16 = mul(x = var_435_cast_fp16, y = var_436_to_fp16)[name = tensor("op_437_cast_fp16")]; + tensor var_438 = const()[name = tensor("op_438"), val = tensor([1, 8, 64, -1])]; + tensor var_439_cast_fp16 = reshape(shape = var_438, x = key_7_cast_fp16)[name = tensor("op_439_cast_fp16")]; + tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; + tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; + tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_437_cast_fp16, y = var_439_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_286, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 8, 64, -1])]; + tensor var_444_cast_fp16 = reshape(shape = var_443, x = value_7_cast_fp16)[name = tensor("op_444_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_444_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 512, 1, -1])]; + tensor input_13_cast_fp16 = reshape(shape = var_447, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor var_451 = const()[name = tensor("op_451"), val = tensor([1, 1])]; + tensor var_453 = const()[name = tensor("op_453"), val = tensor([1, 1])]; + tensor obj_25_pad_type_0 = const()[name = tensor("obj_25_pad_type_0"), val = tensor("custom")]; + tensor obj_25_pad_0 = const()[name = tensor("obj_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65658688)))]; + tensor layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66183040)))]; + tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_453, groups = var_293, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = var_451, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor var_459 = const()[name = tensor("op_459"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_459, keep_dims = var_294, x = inputs_11_cast_fp16)[name = tensor("channels_mean_11_cast_fp16")]; + tensor zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor("zero_mean_11_cast_fp16")]; + tensor zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor("zero_mean_sq_11_cast_fp16")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1])]; + tensor var_464_cast_fp16 = reduce_mean(axes = var_463, keep_dims = var_294, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor var_465_to_fp16 = const()[name = tensor("op_465_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_466_cast_fp16 = add(x = var_464_cast_fp16, y = var_465_to_fp16)[name = tensor("op_466_cast_fp16")]; + tensor denom_11_epsilon_0_to_fp16 = const()[name = tensor("denom_11_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_466_cast_fp16)[name = tensor("denom_11_cast_fp16")]; + tensor out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_15_gamma_0_to_fp16 = const()[name = tensor("input_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66184128)))]; + tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66185216)))]; + tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor var_477 = const()[name = tensor("op_477"), val = tensor([1, 1])]; + tensor var_479 = const()[name = tensor("op_479"), val = tensor([1, 1])]; + tensor input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("custom")]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66186304)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68283520)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_479, groups = var_293, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_477, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor var_485 = const()[name = tensor("op_485"), val = tensor([1, 1])]; + tensor var_487 = const()[name = tensor("op_487"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc2_weight_to_fp16 = const()[name = tensor("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68287680)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70384896)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_487, groups = var_293, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_485, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor(3)]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(1)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(true)]; + tensor var_520 = const()[name = tensor("op_520"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_520, keep_dims = var_508, x = inputs_13_cast_fp16)[name = tensor("channels_mean_13_cast_fp16")]; + tensor zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor("zero_mean_13_cast_fp16")]; + tensor zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor("zero_mean_sq_13_cast_fp16")]; + tensor var_524 = const()[name = tensor("op_524"), val = tensor([1])]; + tensor var_525_cast_fp16 = reduce_mean(axes = var_524, keep_dims = var_508, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_525_cast_fp16")]; + tensor var_526_to_fp16 = const()[name = tensor("op_526_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_527_cast_fp16 = add(x = var_525_cast_fp16, y = var_526_to_fp16)[name = tensor("op_527_cast_fp16")]; + tensor denom_13_epsilon_0_to_fp16 = const()[name = tensor("denom_13_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0_to_fp16, x = var_527_cast_fp16)[name = tensor("denom_13_cast_fp16")]; + tensor out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70385984)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70387072)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor var_542 = const()[name = tensor("op_542"), val = tensor([1, 1])]; + tensor var_544 = const()[name = tensor("op_544"), val = tensor([1, 1])]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("custom")]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70388160)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70912512)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = var_544, groups = var_507, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = var_542, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_548 = const()[name = tensor("op_548"), val = tensor([1, 1])]; + tensor var_550 = const()[name = tensor("op_550"), val = tensor([1, 1])]; + tensor current_key_5_pad_type_0 = const()[name = tensor("current_key_5_pad_type_0"), val = tensor("custom")]; + tensor current_key_5_pad_0 = const()[name = tensor("current_key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70913600)))]; + tensor current_key_5_cast_fp16 = conv(dilations = var_550, groups = var_507, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = var_548, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor var_555 = const()[name = tensor("op_555"), val = tensor([1, 1])]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1])]; + tensor current_value_5_pad_type_0 = const()[name = tensor("current_value_5_pad_type_0"), val = tensor("custom")]; + tensor current_value_5_pad_0 = const()[name = tensor("current_value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71437952)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71962304)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = var_557, groups = var_507, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = var_555, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor var_566_cast_fp16 = mul(x = var_51_cast_fp16_2, y = var_137_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_564_cast_fp16, y = var_566_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_568_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor var_570_cast_fp16 = mul(x = var_60_cast_fp16_2, y = var_137_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_568_cast_fp16, y = var_570_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 8, 64, -1])]; + tensor var_574_cast_fp16 = reshape(shape = var_573, x = query_9_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor var_575_to_fp16 = const()[name = tensor("op_575_to_fp16"), val = tensor(0x1p-3)]; + tensor var_576_cast_fp16 = mul(x = var_574_cast_fp16, y = var_575_to_fp16)[name = tensor("op_576_cast_fp16")]; + tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 8, 64, -1])]; + tensor var_578_cast_fp16 = reshape(shape = var_577, x = key_9_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_576_cast_fp16, y = var_578_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_155_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_586_cast_fp16 = softmax(axis = var_500, x = mh_w_15_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor var_587 = const()[name = tensor("op_587"), val = tensor([1, 8, 64, -1])]; + tensor var_588_cast_fp16 = reshape(shape = var_587, x = value_9_cast_fp16)[name = tensor("op_588_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_588_cast_fp16, y = var_586_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 512, 1, -1])]; + tensor input_21_cast_fp16 = reshape(shape = var_591, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1])]; + tensor var_597 = const()[name = tensor("op_597"), val = tensor([1, 1])]; + tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("custom")]; + tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71963392)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72487744)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = var_597, groups = var_507, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = var_595, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_607, keep_dims = var_508, x = inputs_15_cast_fp16)[name = tensor("channels_mean_15_cast_fp16")]; + tensor zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor("zero_mean_15_cast_fp16")]; + tensor zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor("zero_mean_sq_15_cast_fp16")]; + tensor var_611 = const()[name = tensor("op_611"), val = tensor([1])]; + tensor var_612_cast_fp16 = reduce_mean(axes = var_611, keep_dims = var_508, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_612_cast_fp16")]; + tensor var_613_to_fp16 = const()[name = tensor("op_613_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_614_cast_fp16 = add(x = var_612_cast_fp16, y = var_613_to_fp16)[name = tensor("op_614_cast_fp16")]; + tensor denom_15_epsilon_0_to_fp16 = const()[name = tensor("denom_15_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0_to_fp16, x = var_614_cast_fp16)[name = tensor("denom_15_cast_fp16")]; + tensor out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72488832)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72489920)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 1])]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor([1, 1])]; + tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("custom")]; + tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72491008)))]; + tensor layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73015360)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = var_631, groups = var_507, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = var_629, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([1, 1])]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([1, 1])]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("custom")]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73016448)))]; + tensor key_11_cast_fp16 = conv(dilations = var_637, groups = var_507, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = var_635, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; + tensor var_642 = const()[name = tensor("op_642"), val = tensor([1, 1])]; + tensor var_644 = const()[name = tensor("op_644"), val = tensor([1, 1])]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("custom")]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73540800)))]; + tensor layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(74065152)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = var_644, groups = var_507, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = var_642, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor([1, 8, 64, -1])]; + tensor var_649_cast_fp16 = reshape(shape = var_648, x = query_11_cast_fp16)[name = tensor("op_649_cast_fp16")]; + tensor var_650_to_fp16 = const()[name = tensor("op_650_to_fp16"), val = tensor(0x1p-3)]; + tensor var_651_cast_fp16 = mul(x = var_649_cast_fp16, y = var_650_to_fp16)[name = tensor("op_651_cast_fp16")]; + tensor var_652 = const()[name = tensor("op_652"), val = tensor([1, 8, 64, -1])]; + tensor var_653_cast_fp16 = reshape(shape = var_652, x = key_11_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_651_cast_fp16, y = var_653_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_500, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_657 = const()[name = tensor("op_657"), val = tensor([1, 8, 64, -1])]; + tensor var_658_cast_fp16 = reshape(shape = var_657, x = value_11_cast_fp16)[name = tensor("op_658_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_658_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 512, 1, -1])]; + tensor input_23_cast_fp16 = reshape(shape = var_661, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_665 = const()[name = tensor("op_665"), val = tensor([1, 1])]; + tensor var_667 = const()[name = tensor("op_667"), val = tensor([1, 1])]; + tensor obj_39_pad_type_0 = const()[name = tensor("obj_39_pad_type_0"), val = tensor("custom")]; + tensor obj_39_pad_0 = const()[name = tensor("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(74066240)))]; + tensor layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(74590592)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = var_667, groups = var_507, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = var_665, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_673 = const()[name = tensor("op_673"), val = tensor([1])]; + tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_673, keep_dims = var_508, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; + tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; + tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1])]; + tensor var_678_cast_fp16 = reduce_mean(axes = var_677, keep_dims = var_508, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor var_679_to_fp16 = const()[name = tensor("op_679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_680_cast_fp16 = add(x = var_678_cast_fp16, y = var_679_to_fp16)[name = tensor("op_680_cast_fp16")]; + tensor denom_17_epsilon_0_to_fp16 = const()[name = tensor("denom_17_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0_to_fp16, x = var_680_cast_fp16)[name = tensor("denom_17_cast_fp16")]; + tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_25_gamma_0_to_fp16 = const()[name = tensor("input_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(74591680)))]; + tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(74592768)))]; + tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 1])]; + tensor var_693 = const()[name = tensor("op_693"), val = tensor([1, 1])]; + tensor input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("custom")]; + tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(74593856)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76691072)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = var_693, groups = var_507, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = var_691, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("EXACT")]; + tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 1])]; + tensor var_701 = const()[name = tensor("op_701"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc2_weight_to_fp16 = const()[name = tensor("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76695232)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78792448)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = var_701, groups = var_507, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = var_699, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_714 = const()[name = tensor("op_714"), val = tensor(3)]; + tensor var_721 = const()[name = tensor("op_721"), val = tensor(1)]; + tensor var_722 = const()[name = tensor("op_722"), val = tensor(true)]; + tensor var_734 = const()[name = tensor("op_734"), val = tensor([1])]; + tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_734, keep_dims = var_722, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; + tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; + tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; + tensor var_738 = const()[name = tensor("op_738"), val = tensor([1])]; + tensor var_739_cast_fp16 = reduce_mean(axes = var_738, keep_dims = var_722, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_741_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor denom_19_epsilon_0_to_fp16 = const()[name = tensor("denom_19_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0_to_fp16, x = var_741_cast_fp16)[name = tensor("denom_19_cast_fp16")]; + tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78793536)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78794624)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor var_756 = const()[name = tensor("op_756"), val = tensor([1, 1])]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("custom")]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78795712)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79320064)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = var_758, groups = var_721, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = var_756, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_762 = const()[name = tensor("op_762"), val = tensor([1, 1])]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor([1, 1])]; + tensor current_key_7_pad_type_0 = const()[name = tensor("current_key_7_pad_type_0"), val = tensor("custom")]; + tensor current_key_7_pad_0 = const()[name = tensor("current_key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79321152)))]; + tensor current_key_7_cast_fp16 = conv(dilations = var_764, groups = var_721, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = var_762, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1])]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1])]; + tensor current_value_7_pad_type_0 = const()[name = tensor("current_value_7_pad_type_0"), val = tensor("custom")]; + tensor current_value_7_pad_0 = const()[name = tensor("current_value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79845504)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80369856)))]; + tensor current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = var_771, groups = var_721, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = var_769, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; + tensor var_778_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_780_cast_fp16 = mul(x = var_51_cast_fp16_3, y = var_137_cast_fp16)[name = tensor("op_780_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_778_cast_fp16, y = var_780_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_782_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_782_cast_fp16")]; + tensor var_784_cast_fp16 = mul(x = var_60_cast_fp16_3, y = var_137_cast_fp16)[name = tensor("op_784_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_782_cast_fp16, y = var_784_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1, 8, 64, -1])]; + tensor var_788_cast_fp16 = reshape(shape = var_787, x = query_13_cast_fp16)[name = tensor("op_788_cast_fp16")]; + tensor var_789_to_fp16 = const()[name = tensor("op_789_to_fp16"), val = tensor(0x1p-3)]; + tensor var_790_cast_fp16 = mul(x = var_788_cast_fp16, y = var_789_to_fp16)[name = tensor("op_790_cast_fp16")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor([1, 8, 64, -1])]; + tensor var_792_cast_fp16 = reshape(shape = var_791, x = key_13_cast_fp16)[name = tensor("op_792_cast_fp16")]; + tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; + tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_790_cast_fp16, y = var_792_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_155_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_800_cast_fp16 = softmax(axis = var_714, x = mh_w_21_cast_fp16)[name = tensor("op_800_cast_fp16")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 8, 64, -1])]; + tensor var_802_cast_fp16 = reshape(shape = var_801, x = value_13_cast_fp16)[name = tensor("op_802_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_802_cast_fp16, y = var_800_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 512, 1, -1])]; + tensor input_31_cast_fp16 = reshape(shape = var_805, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_809 = const()[name = tensor("op_809"), val = tensor([1, 1])]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor([1, 1])]; + tensor obj_49_pad_type_0 = const()[name = tensor("obj_49_pad_type_0"), val = tensor("custom")]; + tensor obj_49_pad_0 = const()[name = tensor("obj_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80370944)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80895296)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = var_811, groups = var_721, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = var_809, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_821 = const()[name = tensor("op_821"), val = tensor([1])]; + tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_821, keep_dims = var_722, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; + tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; + tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1])]; + tensor var_826_cast_fp16 = reduce_mean(axes = var_825, keep_dims = var_722, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_827_to_fp16 = const()[name = tensor("op_827_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_828_cast_fp16 = add(x = var_826_cast_fp16, y = var_827_to_fp16)[name = tensor("op_828_cast_fp16")]; + tensor denom_21_epsilon_0_to_fp16 = const()[name = tensor("denom_21_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0_to_fp16, x = var_828_cast_fp16)[name = tensor("denom_21_cast_fp16")]; + tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80896384)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80897472)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 1])]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor([1, 1])]; + tensor query_15_pad_type_0 = const()[name = tensor("query_15_pad_type_0"), val = tensor("custom")]; + tensor query_15_pad_0 = const()[name = tensor("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80898560)))]; + tensor layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81422912)))]; + tensor query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = var_845, groups = var_721, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = var_843, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor var_849 = const()[name = tensor("op_849"), val = tensor([1, 1])]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 1])]; + tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("custom")]; + tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81424000)))]; + tensor key_15_cast_fp16 = conv(dilations = var_851, groups = var_721, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = var_849, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_15_cast_fp16")]; + tensor var_856 = const()[name = tensor("op_856"), val = tensor([1, 1])]; + tensor var_858 = const()[name = tensor("op_858"), val = tensor([1, 1])]; + tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("custom")]; + tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81948352)))]; + tensor layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82472704)))]; + tensor value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = var_858, groups = var_721, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = var_856, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_15_cast_fp16")]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 8, 64, -1])]; + tensor var_863_cast_fp16 = reshape(shape = var_862, x = query_15_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_to_fp16 = const()[name = tensor("op_864_to_fp16"), val = tensor(0x1p-3)]; + tensor var_865_cast_fp16 = mul(x = var_863_cast_fp16, y = var_864_to_fp16)[name = tensor("op_865_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 8, 64, -1])]; + tensor var_867_cast_fp16 = reshape(shape = var_866, x = key_15_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor mh_w_23_transpose_x_0 = const()[name = tensor("mh_w_23_transpose_x_0"), val = tensor(true)]; + tensor mh_w_23_transpose_y_0 = const()[name = tensor("mh_w_23_transpose_y_0"), val = tensor(false)]; + tensor mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_865_cast_fp16, y = var_867_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_714, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_871 = const()[name = tensor("op_871"), val = tensor([1, 8, 64, -1])]; + tensor var_872_cast_fp16 = reshape(shape = var_871, x = value_15_cast_fp16)[name = tensor("op_872_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_872_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 512, 1, -1])]; + tensor input_33_cast_fp16 = reshape(shape = var_875, x = attn_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor var_879 = const()[name = tensor("op_879"), val = tensor([1, 1])]; + tensor var_881 = const()[name = tensor("op_881"), val = tensor([1, 1])]; + tensor obj_53_pad_type_0 = const()[name = tensor("obj_53_pad_type_0"), val = tensor("custom")]; + tensor obj_53_pad_0 = const()[name = tensor("obj_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82473792)))]; + tensor layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82998144)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = var_881, groups = var_721, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = var_879, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor var_890 = const()[name = tensor("op_890"), val = tensor([1])]; + tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_890, keep_dims = var_722, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; + tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; + tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; + tensor var_894 = const()[name = tensor("op_894"), val = tensor([1])]; + tensor var_895_cast_fp16 = reduce_mean(axes = var_894, keep_dims = var_722, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_895_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_897_cast_fp16 = add(x = var_895_cast_fp16, y = var_896_to_fp16)[name = tensor("op_897_cast_fp16")]; + tensor denom_23_epsilon_0_to_fp16 = const()[name = tensor("denom_23_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0_to_fp16, x = var_897_cast_fp16)[name = tensor("denom_23_cast_fp16")]; + tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82999232)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83000320)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor var_908 = const()[name = tensor("op_908"), val = tensor([1, 1])]; + tensor var_910 = const()[name = tensor("op_910"), val = tensor([1, 1])]; + tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("custom")]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc1_weight_to_fp16 = const()[name = tensor("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83001408)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85098624)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = var_910, groups = var_721, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = var_908, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_mode_0 = const()[name = tensor("input_39_mode_0"), val = tensor("EXACT")]; + tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 1])]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc2_weight_to_fp16 = const()[name = tensor("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85102784)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87200000)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = var_918, groups = var_721, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = var_916, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_932 = const()[name = tensor("op_932"), val = tensor(3)]; + tensor var_939 = const()[name = tensor("op_939"), val = tensor(1)]; + tensor var_940 = const()[name = tensor("op_940"), val = tensor(true)]; + tensor var_952 = const()[name = tensor("op_952"), val = tensor([1])]; + tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_952, keep_dims = var_940, x = inputs_25_cast_fp16)[name = tensor("channels_mean_25_cast_fp16")]; + tensor zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor("zero_mean_25_cast_fp16")]; + tensor zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor("zero_mean_sq_25_cast_fp16")]; + tensor var_956 = const()[name = tensor("op_956"), val = tensor([1])]; + tensor var_957_cast_fp16 = reduce_mean(axes = var_956, keep_dims = var_940, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor var_958_to_fp16 = const()[name = tensor("op_958_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_959_cast_fp16 = add(x = var_957_cast_fp16, y = var_958_to_fp16)[name = tensor("op_959_cast_fp16")]; + tensor denom_25_epsilon_0_to_fp16 = const()[name = tensor("denom_25_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0_to_fp16, x = var_959_cast_fp16)[name = tensor("denom_25_cast_fp16")]; + tensor out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87201088)))]; + tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87202176)))]; + tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor var_974 = const()[name = tensor("op_974"), val = tensor([1, 1])]; + tensor var_976 = const()[name = tensor("op_976"), val = tensor([1, 1])]; + tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("custom")]; + tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87203264)))]; + tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87727616)))]; + tensor query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = var_976, groups = var_939, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = var_974, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor var_980 = const()[name = tensor("op_980"), val = tensor([1, 1])]; + tensor var_982 = const()[name = tensor("op_982"), val = tensor([1, 1])]; + tensor current_key_9_pad_type_0 = const()[name = tensor("current_key_9_pad_type_0"), val = tensor("custom")]; + tensor current_key_9_pad_0 = const()[name = tensor("current_key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87728704)))]; + tensor current_key_9_cast_fp16 = conv(dilations = var_982, groups = var_939, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = var_980, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor([1, 1])]; + tensor var_989 = const()[name = tensor("op_989"), val = tensor([1, 1])]; + tensor current_value_9_pad_type_0 = const()[name = tensor("current_value_9_pad_type_0"), val = tensor("custom")]; + tensor current_value_9_pad_0 = const()[name = tensor("current_value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88253056)))]; + tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88777408)))]; + tensor current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = var_989, groups = var_939, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = var_987, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor var_998_cast_fp16 = mul(x = var_51_cast_fp16_4, y = var_137_cast_fp16)[name = tensor("op_998_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_996_cast_fp16, y = var_998_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1000_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_1000_cast_fp16")]; + tensor var_1002_cast_fp16 = mul(x = var_60_cast_fp16_4, y = var_137_cast_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_1000_cast_fp16, y = var_1002_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1005 = const()[name = tensor("op_1005"), val = tensor([1, 8, 64, -1])]; + tensor var_1006_cast_fp16 = reshape(shape = var_1005, x = query_17_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor var_1007_to_fp16 = const()[name = tensor("op_1007_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1008_cast_fp16 = mul(x = var_1006_cast_fp16, y = var_1007_to_fp16)[name = tensor("op_1008_cast_fp16")]; + tensor var_1009 = const()[name = tensor("op_1009"), val = tensor([1, 8, 64, -1])]; + tensor var_1010_cast_fp16 = reshape(shape = var_1009, x = key_17_cast_fp16)[name = tensor("op_1010_cast_fp16")]; + tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; + tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; + tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1008_cast_fp16, y = var_1010_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_155_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1018_cast_fp16 = softmax(axis = var_932, x = mh_w_27_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([1, 8, 64, -1])]; + tensor var_1020_cast_fp16 = reshape(shape = var_1019, x = value_17_cast_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1020_cast_fp16, y = var_1018_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 512, 1, -1])]; + tensor input_41_cast_fp16 = reshape(shape = var_1023, x = attn_17_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([1, 1])]; + tensor var_1029 = const()[name = tensor("op_1029"), val = tensor([1, 1])]; + tensor obj_63_pad_type_0 = const()[name = tensor("obj_63_pad_type_0"), val = tensor("custom")]; + tensor obj_63_pad_0 = const()[name = tensor("obj_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88778496)))]; + tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89302848)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = var_1029, groups = var_939, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = var_1027, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor var_1039 = const()[name = tensor("op_1039"), val = tensor([1])]; + tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1039, keep_dims = var_940, x = inputs_27_cast_fp16)[name = tensor("channels_mean_27_cast_fp16")]; + tensor zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor("zero_mean_27_cast_fp16")]; + tensor zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor("zero_mean_sq_27_cast_fp16")]; + tensor var_1043 = const()[name = tensor("op_1043"), val = tensor([1])]; + tensor var_1044_cast_fp16 = reduce_mean(axes = var_1043, keep_dims = var_940, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1044_cast_fp16")]; + tensor var_1045_to_fp16 = const()[name = tensor("op_1045_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1046_cast_fp16 = add(x = var_1044_cast_fp16, y = var_1045_to_fp16)[name = tensor("op_1046_cast_fp16")]; + tensor denom_27_epsilon_0_to_fp16 = const()[name = tensor("denom_27_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0_to_fp16, x = var_1046_cast_fp16)[name = tensor("denom_27_cast_fp16")]; + tensor out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89303936)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89305024)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor var_1061 = const()[name = tensor("op_1061"), val = tensor([1, 1])]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1])]; + tensor query_19_pad_type_0 = const()[name = tensor("query_19_pad_type_0"), val = tensor("custom")]; + tensor query_19_pad_0 = const()[name = tensor("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89306112)))]; + tensor layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89830464)))]; + tensor query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = var_1063, groups = var_939, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = var_1061, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor var_1067 = const()[name = tensor("op_1067"), val = tensor([1, 1])]; + tensor var_1069 = const()[name = tensor("op_1069"), val = tensor([1, 1])]; + tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("custom")]; + tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89831552)))]; + tensor key_19_cast_fp16 = conv(dilations = var_1069, groups = var_939, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = var_1067, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_19_cast_fp16")]; + tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([1, 1])]; + tensor var_1076 = const()[name = tensor("op_1076"), val = tensor([1, 1])]; + tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("custom")]; + tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90355904)))]; + tensor layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90880256)))]; + tensor value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = var_1076, groups = var_939, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = var_1074, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_19_cast_fp16")]; + tensor var_1080 = const()[name = tensor("op_1080"), val = tensor([1, 8, 64, -1])]; + tensor var_1081_cast_fp16 = reshape(shape = var_1080, x = query_19_cast_fp16)[name = tensor("op_1081_cast_fp16")]; + tensor var_1082_to_fp16 = const()[name = tensor("op_1082_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1083_cast_fp16 = mul(x = var_1081_cast_fp16, y = var_1082_to_fp16)[name = tensor("op_1083_cast_fp16")]; + tensor var_1084 = const()[name = tensor("op_1084"), val = tensor([1, 8, 64, -1])]; + tensor var_1085_cast_fp16 = reshape(shape = var_1084, x = key_19_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; + tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; + tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1083_cast_fp16, y = var_1085_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_932, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor([1, 8, 64, -1])]; + tensor var_1090_cast_fp16 = reshape(shape = var_1089, x = value_19_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1090_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([1, 512, 1, -1])]; + tensor input_43_cast_fp16 = reshape(shape = var_1093, x = attn_19_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([1, 1])]; + tensor var_1099 = const()[name = tensor("op_1099"), val = tensor([1, 1])]; + tensor obj_67_pad_type_0 = const()[name = tensor("obj_67_pad_type_0"), val = tensor("custom")]; + tensor obj_67_pad_0 = const()[name = tensor("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90881344)))]; + tensor layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91405696)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = var_1099, groups = var_939, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = var_1097, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1108 = const()[name = tensor("op_1108"), val = tensor([1])]; + tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_1108, keep_dims = var_940, x = inputs_29_cast_fp16)[name = tensor("channels_mean_29_cast_fp16")]; + tensor zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor("zero_mean_29_cast_fp16")]; + tensor zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor("zero_mean_sq_29_cast_fp16")]; + tensor var_1112 = const()[name = tensor("op_1112"), val = tensor([1])]; + tensor var_1113_cast_fp16 = reduce_mean(axes = var_1112, keep_dims = var_940, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor var_1114_to_fp16 = const()[name = tensor("op_1114_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1115_cast_fp16 = add(x = var_1113_cast_fp16, y = var_1114_to_fp16)[name = tensor("op_1115_cast_fp16")]; + tensor denom_29_epsilon_0_to_fp16 = const()[name = tensor("denom_29_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0_to_fp16, x = var_1115_cast_fp16)[name = tensor("denom_29_cast_fp16")]; + tensor out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor input_45_gamma_0_to_fp16 = const()[name = tensor("input_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91406784)))]; + tensor input_45_beta_0_to_fp16 = const()[name = tensor("input_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91407872)))]; + tensor input_45_epsilon_0_to_fp16 = const()[name = tensor("input_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor var_1126 = const()[name = tensor("op_1126"), val = tensor([1, 1])]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1])]; + tensor input_47_pad_type_0 = const()[name = tensor("input_47_pad_type_0"), val = tensor("custom")]; + tensor input_47_pad_0 = const()[name = tensor("input_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc1_weight_to_fp16 = const()[name = tensor("layers_4_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91408960)))]; + tensor layers_4_fc1_bias_to_fp16 = const()[name = tensor("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93506176)))]; + tensor input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = var_1128, groups = var_939, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = var_1126, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor var_1134 = const()[name = tensor("op_1134"), val = tensor([1, 1])]; + tensor var_1136 = const()[name = tensor("op_1136"), val = tensor([1, 1])]; + tensor hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc2_weight_to_fp16 = const()[name = tensor("layers_4_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93510336)))]; + tensor layers_4_fc2_bias_to_fp16 = const()[name = tensor("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95607552)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = var_1136, groups = var_939, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = var_1134, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_1150 = const()[name = tensor("op_1150"), val = tensor(3)]; + tensor var_1157 = const()[name = tensor("op_1157"), val = tensor(1)]; + tensor var_1158 = const()[name = tensor("op_1158"), val = tensor(true)]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([1])]; + tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_1170, keep_dims = var_1158, x = inputs_31_cast_fp16)[name = tensor("channels_mean_31_cast_fp16")]; + tensor zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor("zero_mean_31_cast_fp16")]; + tensor zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor("zero_mean_sq_31_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([1])]; + tensor var_1175_cast_fp16 = reduce_mean(axes = var_1174, keep_dims = var_1158, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_1175_cast_fp16")]; + tensor var_1176_to_fp16 = const()[name = tensor("op_1176_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1177_cast_fp16 = add(x = var_1175_cast_fp16, y = var_1176_to_fp16)[name = tensor("op_1177_cast_fp16")]; + tensor denom_31_epsilon_0_to_fp16 = const()[name = tensor("denom_31_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0_to_fp16, x = var_1177_cast_fp16)[name = tensor("denom_31_cast_fp16")]; + tensor out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor obj_71_gamma_0_to_fp16 = const()[name = tensor("obj_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95608640)))]; + tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95609728)))]; + tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 1])]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([1, 1])]; + tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("custom")]; + tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95610816)))]; + tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96135168)))]; + tensor query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = var_1194, groups = var_1157, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = var_1192, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([1, 1])]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([1, 1])]; + tensor current_key_pad_type_0 = const()[name = tensor("current_key_pad_type_0"), val = tensor("custom")]; + tensor current_key_pad_0 = const()[name = tensor("current_key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96136256)))]; + tensor current_key_cast_fp16 = conv(dilations = var_1200, groups = var_1157, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_1198, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([1, 1])]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([1, 1])]; + tensor current_value_pad_type_0 = const()[name = tensor("current_value_pad_type_0"), val = tensor("custom")]; + tensor current_value_pad_0 = const()[name = tensor("current_value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96660608)))]; + tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97184960)))]; + tensor current_value_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = var_1207, groups = var_1157, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_1205, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_1214_cast_fp16 = mul(x = current_key_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_1214_cast_fp16")]; + tensor var_1216_cast_fp16 = mul(x = var_51_cast_fp16_5, y = var_137_cast_fp16)[name = tensor("op_1216_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_1214_cast_fp16, y = var_1216_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1218_cast_fp16 = mul(x = current_value_cast_fp16, y = var_134_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor var_1220_cast_fp16 = mul(x = var_60_cast_fp16_5, y = var_137_cast_fp16)[name = tensor("op_1220_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_1218_cast_fp16, y = var_1220_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 8, 64, -1])]; + tensor var_1224_cast_fp16 = reshape(shape = var_1223, x = query_21_cast_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor var_1225_to_fp16 = const()[name = tensor("op_1225_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1226_cast_fp16 = mul(x = var_1224_cast_fp16, y = var_1225_to_fp16)[name = tensor("op_1226_cast_fp16")]; + tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([1, 8, 64, -1])]; + tensor var_1228_cast_fp16 = reshape(shape = var_1227, x = key_21_cast_fp16)[name = tensor("op_1228_cast_fp16")]; + tensor mh_w_31_transpose_x_0 = const()[name = tensor("mh_w_31_transpose_x_0"), val = tensor(true)]; + tensor mh_w_31_transpose_y_0 = const()[name = tensor("mh_w_31_transpose_y_0"), val = tensor(false)]; + tensor mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_1226_cast_fp16, y = var_1228_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_155_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_1236_cast_fp16 = softmax(axis = var_1150, x = mh_w_33_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 8, 64, -1])]; + tensor var_1238_cast_fp16 = reshape(shape = var_1237, x = value_21_cast_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1238_cast_fp16, y = var_1236_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([1, 512, 1, -1])]; + tensor input_51_cast_fp16 = reshape(shape = var_1241, x = attn_21_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([1, 1])]; + tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([1, 1])]; + tensor obj_77_pad_type_0 = const()[name = tensor("obj_77_pad_type_0"), val = tensor("custom")]; + tensor obj_77_pad_0 = const()[name = tensor("obj_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97186048)))]; + tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97710400)))]; + tensor obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = var_1247, groups = var_1157, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = var_1245, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([1])]; + tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_1257, keep_dims = var_1158, x = inputs_33_cast_fp16)[name = tensor("channels_mean_33_cast_fp16")]; + tensor zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor("zero_mean_33_cast_fp16")]; + tensor zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor("zero_mean_sq_33_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([1])]; + tensor var_1262_cast_fp16 = reduce_mean(axes = var_1261, keep_dims = var_1158, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1263_to_fp16 = const()[name = tensor("op_1263_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1264_cast_fp16 = add(x = var_1262_cast_fp16, y = var_1263_to_fp16)[name = tensor("op_1264_cast_fp16")]; + tensor denom_33_epsilon_0_to_fp16 = const()[name = tensor("denom_33_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0_to_fp16, x = var_1264_cast_fp16)[name = tensor("denom_33_cast_fp16")]; + tensor out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_79_gamma_0_to_fp16 = const()[name = tensor("obj_79_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97711488)))]; + tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97712576)))]; + tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 1])]; + tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 1])]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("custom")]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(97713664)))]; + tensor layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98238016)))]; + tensor query_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = var_1281, groups = var_1157, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_1279, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([1, 1])]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([1, 1])]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("custom")]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98239104)))]; + tensor key_cast_fp16 = conv(dilations = var_1287, groups = var_1157, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_1285, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; + tensor var_1292 = const()[name = tensor("op_1292"), val = tensor([1, 1])]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([1, 1])]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("custom")]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98763456)))]; + tensor layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99287808)))]; + tensor value_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = var_1294, groups = var_1157, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_1292, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 8, 64, -1])]; + tensor var_1299_cast_fp16 = reshape(shape = var_1298, x = query_cast_fp16)[name = tensor("op_1299_cast_fp16")]; + tensor var_1300_to_fp16 = const()[name = tensor("op_1300_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1301_cast_fp16 = mul(x = var_1299_cast_fp16, y = var_1300_to_fp16)[name = tensor("op_1301_cast_fp16")]; + tensor var_1302 = const()[name = tensor("op_1302"), val = tensor([1, 8, 64, -1])]; + tensor var_1303_cast_fp16 = reshape(shape = var_1302, x = key_cast_fp16)[name = tensor("op_1303_cast_fp16")]; + tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; + tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_1301_cast_fp16, y = var_1303_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_1150, x = mh_w_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_1307 = const()[name = tensor("op_1307"), val = tensor([1, 8, 64, -1])]; + tensor var_1308_cast_fp16 = reshape(shape = var_1307, x = value_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_1308_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_1311 = const()[name = tensor("op_1311"), val = tensor([1, 512, 1, -1])]; + tensor input_53_cast_fp16 = reshape(shape = var_1311, x = attn_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor var_1315 = const()[name = tensor("op_1315"), val = tensor([1, 1])]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([1, 1])]; + tensor obj_81_pad_type_0 = const()[name = tensor("obj_81_pad_type_0"), val = tensor("custom")]; + tensor obj_81_pad_0 = const()[name = tensor("obj_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99288896)))]; + tensor layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99813248)))]; + tensor obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = var_1317, groups = var_1157, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = var_1315, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([1])]; + tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_1326, keep_dims = var_1158, x = inputs_35_cast_fp16)[name = tensor("channels_mean_35_cast_fp16")]; + tensor zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor("zero_mean_35_cast_fp16")]; + tensor zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor("zero_mean_sq_35_cast_fp16")]; + tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1])]; + tensor var_1331_cast_fp16 = reduce_mean(axes = var_1330, keep_dims = var_1158, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_1331_cast_fp16")]; + tensor var_1332_to_fp16 = const()[name = tensor("op_1332_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1333_cast_fp16 = add(x = var_1331_cast_fp16, y = var_1332_to_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor denom_35_epsilon_0_to_fp16 = const()[name = tensor("denom_35_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0_to_fp16, x = var_1333_cast_fp16)[name = tensor("denom_35_cast_fp16")]; + tensor out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_55_gamma_0_to_fp16 = const()[name = tensor("input_55_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99814336)))]; + tensor input_55_beta_0_to_fp16 = const()[name = tensor("input_55_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99815424)))]; + tensor input_55_epsilon_0_to_fp16 = const()[name = tensor("input_55_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_55_cast_fp16 = batch_norm(beta = input_55_beta_0_to_fp16, epsilon = input_55_epsilon_0_to_fp16, gamma = input_55_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([1, 1])]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([1, 1])]; + tensor input_57_pad_type_0 = const()[name = tensor("input_57_pad_type_0"), val = tensor("custom")]; + tensor input_57_pad_0 = const()[name = tensor("input_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc1_weight_to_fp16 = const()[name = tensor("layers_5_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99816512)))]; + tensor layers_5_fc1_bias_to_fp16 = const()[name = tensor("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101913728)))]; + tensor input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = var_1346, groups = var_1157, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = var_1344, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_57_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_1352 = const()[name = tensor("op_1352"), val = tensor([1, 1])]; + tensor var_1354 = const()[name = tensor("op_1354"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("custom")]; + tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc2_weight_to_fp16 = const()[name = tensor("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101917888)))]; + tensor layers_5_fc2_bias_to_fp16 = const()[name = tensor("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104015104)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = var_1354, groups = var_1157, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = var_1352, weight = layers_5_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor var_1365 = const()[name = tensor("op_1365"), val = tensor(true)]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_1369, keep_dims = var_1365, x = inputs_cast_fp16)[name = tensor("channels_mean_cast_fp16")]; + tensor zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor("zero_mean_cast_fp16")]; + tensor zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor("zero_mean_sq_cast_fp16")]; + tensor var_1373 = const()[name = tensor("op_1373"), val = tensor([1])]; + tensor var_1374_cast_fp16 = reduce_mean(axes = var_1373, keep_dims = var_1365, x = zero_mean_sq_cast_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor var_1375_to_fp16 = const()[name = tensor("op_1375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1376_cast_fp16 = add(x = var_1374_cast_fp16, y = var_1375_to_fp16)[name = tensor("op_1376_cast_fp16")]; + tensor denom_epsilon_0_to_fp16 = const()[name = tensor("denom_epsilon_0_to_fp16"), val = tensor(0x1p-24)]; + tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_1376_cast_fp16)[name = tensor("denom_cast_fp16")]; + tensor out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104016192)))]; + tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104017280)))]; + tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor var_1386_axes_0 = const()[name = tensor("op_1386_axes_0"), val = tensor([2])]; + tensor var_1386_cast_fp16 = squeeze(axes = var_1386_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor var_1389_perm_0 = const()[name = tensor("op_1389_perm_0"), val = tensor([0, 2, 1])]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104018368)))]; + tensor transpose_0 = transpose(perm = var_1389_perm_0, x = var_1386_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor("linear_0_cast_fp16")]; + tensor var_1393 = const()[name = tensor("op_1393"), val = tensor(1)]; + tensor obj_87_interleave_0 = const()[name = tensor("obj_87_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_1393, interleave = obj_87_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_cast_fp16))[name = tensor("obj_87_cast_fp16")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor(1)]; + tensor obj_89_interleave_0 = const()[name = tensor("obj_89_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_1396, interleave = obj_89_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_cast_fp16))[name = tensor("obj_89_cast_fp16")]; + tensor var_1407_begin_0 = const()[name = tensor("op_1407_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1407_end_0 = const()[name = tensor("op_1407_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_1407_end_mask_0 = const()[name = tensor("op_1407_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1407_cast_fp16 = slice_by_index(begin = var_1407_begin_0, end = var_1407_end_0, end_mask = var_1407_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1407_cast_fp16")]; + tensor var_1410_begin_0 = const()[name = tensor("op_1410_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1410_end_0 = const()[name = tensor("op_1410_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1410_end_mask_0 = const()[name = tensor("op_1410_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1410_squeeze_mask_0 = const()[name = tensor("op_1410_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1410_cast_fp16 = slice_by_index(begin = var_1410_begin_0, end = var_1410_end_0, end_mask = var_1410_end_mask_0, squeeze_mask = var_1410_squeeze_mask_0, x = var_1407_cast_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1425_begin_0 = const()[name = tensor("op_1425_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1425_end_0 = const()[name = tensor("op_1425_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_1425_end_mask_0 = const()[name = tensor("op_1425_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1425_cast_fp16 = slice_by_index(begin = var_1425_begin_0, end = var_1425_end_0, end_mask = var_1425_end_mask_0, x = obj_69_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1428_begin_0 = const()[name = tensor("op_1428_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1428_end_0 = const()[name = tensor("op_1428_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1428_end_mask_0 = const()[name = tensor("op_1428_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1428_squeeze_mask_0 = const()[name = tensor("op_1428_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1428_cast_fp16 = slice_by_index(begin = var_1428_begin_0, end = var_1428_end_0, end_mask = var_1428_end_mask_0, squeeze_mask = var_1428_squeeze_mask_0, x = var_1425_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor var_1443_begin_0 = const()[name = tensor("op_1443_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1443_end_0 = const()[name = tensor("op_1443_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_1443_end_mask_0 = const()[name = tensor("op_1443_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1443_cast_fp16 = slice_by_index(begin = var_1443_begin_0, end = var_1443_end_0, end_mask = var_1443_end_mask_0, x = obj_69_cast_fp16)[name = tensor("op_1443_cast_fp16")]; + tensor var_1446_begin_0 = const()[name = tensor("op_1446_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1446_end_0 = const()[name = tensor("op_1446_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1446_end_mask_0 = const()[name = tensor("op_1446_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1446_squeeze_mask_0 = const()[name = tensor("op_1446_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1446_cast_fp16 = slice_by_index(begin = var_1446_begin_0, end = var_1446_end_0, end_mask = var_1446_end_mask_0, squeeze_mask = var_1446_squeeze_mask_0, x = var_1443_cast_fp16)[name = tensor("op_1446_cast_fp16")]; + tensor var_1461_begin_0 = const()[name = tensor("op_1461_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_1461_end_0 = const()[name = tensor("op_1461_end_0"), val = tensor([1, 8, 1, 1500])]; + tensor var_1461_end_mask_0 = const()[name = tensor("op_1461_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1461_cast_fp16 = slice_by_index(begin = var_1461_begin_0, end = var_1461_end_0, end_mask = var_1461_end_mask_0, x = obj_69_cast_fp16)[name = tensor("op_1461_cast_fp16")]; + tensor var_1464_begin_0 = const()[name = tensor("op_1464_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1464_end_0 = const()[name = tensor("op_1464_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1464_end_mask_0 = const()[name = tensor("op_1464_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1464_squeeze_mask_0 = const()[name = tensor("op_1464_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1464_cast_fp16 = slice_by_index(begin = var_1464_begin_0, end = var_1464_end_0, end_mask = var_1464_end_mask_0, squeeze_mask = var_1464_squeeze_mask_0, x = var_1461_cast_fp16)[name = tensor("op_1464_cast_fp16")]; + tensor var_1479_begin_0 = const()[name = tensor("op_1479_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_1479_end_0 = const()[name = tensor("op_1479_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_1479_end_mask_0 = const()[name = tensor("op_1479_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1479_cast_fp16 = slice_by_index(begin = var_1479_begin_0, end = var_1479_end_0, end_mask = var_1479_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor var_1482_begin_0 = const()[name = tensor("op_1482_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1482_end_0 = const()[name = tensor("op_1482_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1482_end_mask_0 = const()[name = tensor("op_1482_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1482_squeeze_mask_0 = const()[name = tensor("op_1482_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1482_cast_fp16 = slice_by_index(begin = var_1482_begin_0, end = var_1482_end_0, end_mask = var_1482_end_mask_0, squeeze_mask = var_1482_squeeze_mask_0, x = var_1479_cast_fp16)[name = tensor("op_1482_cast_fp16")]; + tensor var_1497_begin_0 = const()[name = tensor("op_1497_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_1497_end_0 = const()[name = tensor("op_1497_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_1497_end_mask_0 = const()[name = tensor("op_1497_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1497_cast_fp16 = slice_by_index(begin = var_1497_begin_0, end = var_1497_end_0, end_mask = var_1497_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_1497_cast_fp16")]; + tensor var_1500_begin_0 = const()[name = tensor("op_1500_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1500_end_0 = const()[name = tensor("op_1500_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1500_end_mask_0 = const()[name = tensor("op_1500_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1500_squeeze_mask_0 = const()[name = tensor("op_1500_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1500_cast_fp16 = slice_by_index(begin = var_1500_begin_0, end = var_1500_end_0, end_mask = var_1500_end_mask_0, squeeze_mask = var_1500_squeeze_mask_0, x = var_1497_cast_fp16)[name = tensor("op_1500_cast_fp16")]; + tensor var_1515_begin_0 = const()[name = tensor("op_1515_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1515_end_0 = const()[name = tensor("op_1515_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_1515_end_mask_0 = const()[name = tensor("op_1515_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1515_cast_fp16 = slice_by_index(begin = var_1515_begin_0, end = var_1515_end_0, end_mask = var_1515_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_1515_cast_fp16")]; + tensor var_1518_begin_0 = const()[name = tensor("op_1518_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1518_end_0 = const()[name = tensor("op_1518_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1518_end_mask_0 = const()[name = tensor("op_1518_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1518_squeeze_mask_0 = const()[name = tensor("op_1518_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1518_cast_fp16 = slice_by_index(begin = var_1518_begin_0, end = var_1518_end_0, end_mask = var_1518_end_mask_0, squeeze_mask = var_1518_squeeze_mask_0, x = var_1515_cast_fp16)[name = tensor("op_1518_cast_fp16")]; + tensor var_1533_begin_0 = const()[name = tensor("op_1533_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1533_end_0 = const()[name = tensor("op_1533_end_0"), val = tensor([1, 7, 1, 1500])]; + tensor var_1533_end_mask_0 = const()[name = tensor("op_1533_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1533_cast_fp16 = slice_by_index(begin = var_1533_begin_0, end = var_1533_end_0, end_mask = var_1533_end_mask_0, x = obj_83_cast_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1536_begin_0 = const()[name = tensor("op_1536_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1536_end_0 = const()[name = tensor("op_1536_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1536_end_mask_0 = const()[name = tensor("op_1536_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1536_squeeze_mask_0 = const()[name = tensor("op_1536_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1536_cast_fp16 = slice_by_index(begin = var_1536_begin_0, end = var_1536_end_0, end_mask = var_1536_end_mask_0, squeeze_mask = var_1536_squeeze_mask_0, x = var_1533_cast_fp16)[name = tensor("op_1536_cast_fp16")]; + tensor var_1543 = const()[name = tensor("op_1543"), val = tensor(1)]; + tensor var_1544_interleave_0 = const()[name = tensor("op_1544_interleave_0"), val = tensor(false)]; + tensor var_1544_cast_fp16 = concat(axis = var_1543, interleave = var_1544_interleave_0, values = (var_1410_cast_fp16, var_1428_cast_fp16, var_1446_cast_fp16, var_1464_cast_fp16, var_1482_cast_fp16, var_1500_cast_fp16, var_1518_cast_fp16, var_1536_cast_fp16))[name = tensor("op_1544_cast_fp16")]; + tensor var_1546 = const()[name = tensor("op_1546"), val = tensor([1])]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = var_1546, keep_dims = var_1547, x = var_1544_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file